Extend circuit breaker functionality to support per-model isolation

[kacpersaw]

Summary

Extends the circuit breaker implementation to isolate failures at the model level, not just per-endpoint. This prevents a failing model (e.g., claude-sonnet-4-20250514 being rate-limited) from blocking requests to other models on the same endpoint.

Problem

With endpoint-only circuit breakers, if one model starts failing (e.g., due to rate limiting), all requests to that endpoint get blocked - even for healthy models.

Solution

Circuit breakers are now keyed by endpoint:model tuple instead of just endpoint. Each model gets independent failure tracking and circuit state.

Changes

• Circuit breaker key: Changed from endpoint to endpoint:model
• Metrics labels: Added model label to all circuit breaker metrics
• Architecture: Replaced middleware pattern with Execute() method to access model after request parsing
• Tests: Added per-model isolation tests

Example

/v1/messages:claude-sonnet-4-20250514  → OPEN (rate limited)
/v1/messages:claude-3-5-haiku-20241022 → CLOSED (healthy)

Requests to haiku continue working even when sonnet is blocked.

Files Changed

File	Change
circuitbreaker/circuitbreaker.go	Get(endpoint, model), Execute() method
bridge.go	Pass model to circuit breaker after interceptor creation
metrics/metrics.go	Added model label to CB metrics
circuitbreaker/circuitbreaker_test.go	Per-model isolation tests
circuit_breaker_integration_test.go	Integration test for model isolation